“"fast multipole method" developed for computational physics to rapidly estimate (to arbitrary precision) the conjugate gradient of an error function. (In other words, they tweak the parameters and "get a little better" the next time through the training data.)”